This LoRA kinda sucks.
And that semi-intended pun notwithstanding, I really mean it. I got tired of clicking refresh on Civitai waiting for any explicit NSFW SDXL content, so I created this LoRA with virtually no effort.
I recommend using the Dreamshaper SDXL model and ADetailer (ADetailer is an Automatic1111 extension that uses a face-detection algorithm - as a post process - to identify faces in the image and then regenerate/inpaint them at a higher resolution, ensuring that faces are crisp even in lower resolution images and that Lora likenesses remain sharp/accurate even in shots where the face is farther away from the "camera").
This LoRA mostly generates POV images of people holding an erect penis near their face and in their hands, though also on occasion produces images of people with penises in their mouth, under their butt, or against their genitals. You'll want to play with the weight if you want variety, because it's over trained, and changing the resolutions and orientations between portrait and landscape has a big impact as well.
I spent around 10 minutes grabbing 377 low-quality screenshots from Google Images that were derived from two porn sites, and did zero cropping or editing before training (though I did generated CLIP captions with Kohya SS), and it shows.
For instance, you'll likely want to add a negative for logo/text/watermark, because all of the pics I used had one and I was too lazy to edit it out.
So why even upload this if I think it stinks and I'm too lazy to fix it? Again, there's nothing like this on Civitai for SDXL right now, and there keeps not being anything (Update: I wrote this before I saw https://civarchive.com/models/120628/wts-whatthesuck-sdxl, which just came out today, because my upload speed is super slow - I love that they also have a Santa fellatio pic in their demo images, lmao), which I hope just means that the talented contributors here are busy training and updating their models properly and with care, so until then, maybe someone else can have some fun using this before the good stuff comes out.
Cheers!
Description
FAQ
Comments (14)
Hah, glad you like my Santa ;) Yepp, I was annoyed too that there were just nothing porn related done for SDXL. I hope people will now see the potential and that they will surpass 1.5 easily in no time.
Haha, yeah I guess we'll both be getting coal this year! ;P
It's great to see there is more porn related stuff around now, but my SDXL Lora/Checkpoint collection is still envious of my 1.5 collection, so...we've still got a ways to go yet! :D
Great work on the JustCum Lora too, by the way!
Hi. Can you tell what's the setting you are using to train model? (text from json if you can). And also how much time did it take? I have RTX 4080 and compared to 1.5 lora training is EXTRA slow.
Hi AiTimeIsNow! Sure thing; see the JSON configuration information below. If I recall correctly, this took about 18 hours to train on a 4090, so for a LoRA...yeah, that was EXTRA slow. I probably could have done with out the +1,000 reg images, so removing those would have sped it up quite a bit.
Also FYI, if you're using Automatic1111, you can get the below JSON information from the LoRA UI-card's properties button in the Extra Networks section (the icon for this only appears when you hover your mouse over the LoRa card).
Good luck and cheers!
```
{
"ss_sd_model_name": "sd_xl_base_1.0_0.9vae.safetensors",
"ss_resolution": "(1024, 1024)",
"ss_clip_skip": "None",
"ss_num_train_images": "2639",
"ss_tag_frequency": {
"7_sucking cock": {
"sucking cock": 377
},
"1_woman": {
"woman": 1034
}
},
"ss_training_started_at": "1690717158.1702318",
"ss_optimizer": "transformers.optimization.Adafactor(scale_parameter=False,relative_step=False,warmup_init=False)",
"ss_caption_tag_dropout_rate": "0.0",
"ss_training_comment": "None",
"ss_adaptive_noise_scale": "None",
"ss_seed": "3980802987",
"ss_enable_bucket": "True",
"ss_lr_warmup_steps": "0",
"ss_num_epochs": "10",
"ss_random_crop": "False",
"ss_text_encoder_lr": "0.0004",
"ss_max_bucket_reso": "2048",
"ss_gradient_accumulation_steps": "1",
"ss_total_batch_size": "1",
"ss_scale_weight_norms": "None",
"ss_num_reg_images": "1034",
"ss_lr_scheduler": "cosine",
"ss_unet_lr": "0.0004",
"ss_new_sd_model_hash": "e6bb9ea85bbf7bf6478a7c6d18b71246f22e95d41bcdd80ed40aa212c33cfeff",
"ss_color_aug": "False",
"ss_bucket_info": {
"buckets": {
"0": {
"resolution": [
256,
192
],
"count": 7
},
"1": {
"resolution": [
320,
192
],
"count": 14
},
"2": {
"resolution": [
384,
192
],
"count": 14
},
"3": {
"resolution": [
576,
768
],
"count": 28
},
"4": {
"resolution": [
576,
896
],
"count": 1498
},
"5": {
"resolution": [
640,
320
],
"count": 28
},
"6": {
"resolution": [
768,
1152
],
"count": 21
},
"7": {
"resolution": [
832,
1088
],
"count": 35
},
"8": {
"resolution": [
832,
1216
],
"count": 70
},
"9": {
"resolution": [
896,
576
],
"count": 609
},
"10": {
"resolution": [
896,
640
],
"count": 14
},
"11": {
"resolution": [
1024,
1024
],
"count": 2639
},
"12": {
"resolution": [
1216,
704
],
"count": 7
},
"13": {
"resolution": [
1280,
704
],
"count": 70
},
"14": {
"resolution": [
1344,
768
],
"count": 224
}
},
"mean_img_ar_error": 0.009668937008096769
},
"ss_sd_model_hash": "be9edd61",
"sshs_legacy_hash": "5ad7f880",
"ss_cache_latents": "True",
"ss_network_dropout": "None",
"ss_shuffle_caption": "False",
"ss_v2": "False",
"ss_output_name": "oralsdxl01",
"ss_reg_dataset_dirs": {
"1_woman": {
"n_repeats": 1,
"img_count": 1034
}
},
"ss_keep_tokens": "0",
"ss_learning_rate": "0.0004",
"ss_min_snr_gamma": "None",
"ss_network_alpha": "1.0",
"ss_bucket_no_upscale": "True",
"ss_mixed_precision": "bf16",
"ss_min_bucket_reso": "256",
"ss_steps": "52780",
"ss_lowram": "False",
"ss_caption_dropout_every_n_epochs": "0",
"ss_full_fp16": "False",
"ss_batch_size_per_device": "1",
"ss_multires_noise_discount": "0.3",
"ss_face_crop_aug_range": "None",
"ss_flip_aug": "False",
"ss_noise_offset": "0.0357",
"ss_network_module": "networks.lora",
"ss_max_grad_norm": "1.0",
"ss_caption_dropout_rate": "0.0",
"ss_dataset_dirs": {
"7_sucking cock": {
"n_repeats": 7,
"img_count": 377
}
},
"ss_epoch": "10",
"ss_training_finished_at": "1690784131.3586023",
"sshs_model_hash": "c86a5a03a6184fcd570e6a128815497d622c0c284d029e5a759995411b3b78ae",
"ss_prior_loss_weight": "1.0",
"ss_num_batches_per_epoch": "5278",
"ss_base_model_version": "sdxl_base_v0-9",
"ss_network_dim": "256",
"ss_max_train_steps": "52780",
"ss_max_token_length": "None",
"ss_sd_scripts_commit_hash": "397bf51a8cd36104e52055358e4ffd066c5858df",
"ss_multires_noise_iterations": "None",
"ss_gradient_checkpointing": "True",
"ss_session_id": "4179311610"
}
```
(Also FYI, I had Prettified the JSON for you when I pasted it into my last comment, but saving the comment removed that formatting, so I guess this is the best I can do here...)
@redcreamsoda THANKS!!!
@AiTimeIsNow Welcome! 😄
Your image prompts feature stuff like
ADetailer model: face_yolov8n.pt, ADetailer version: 23.7.11, "cockteaseLoRASDXL: c86a5a03a618", ADetailer mask blur: 4, ADetailer confidence: 0.3, ADetailer dilate/erode: 4, ADetailer inpaint padding: 32, ADetailer denoising strength: 0.4, ADetailer inpaint only masked: True
What is this "ADetailer" thing and how do we replicate it in automatic1111?
Is an extension to automatic fix faces/hands https://github.com/Bing-su/adetailer
Thanks, jhonstn! Yeah, I had linked ADetailer in the description, but I didn't really explain it at all.
ADetailer is a must-have extension for me in SD/Automatic1111, as it uses a face-detection algorithm - as a post process - to identify faces in the image and then regenerate and inpaint them at a higher resolution. This ensures that faces are crisp even in lower resolution images and that Lora likenesses remain sharp/accurate even in shots where the face is farther away from the "camera".
It fixes hands too, reportedly, but I never use it for that, so I can't really speak to that aspect.
Cheers!
Awesome! Hope we're able to figure out smaller XL Loras soon lol, out of space on my PC!
Agreed! I think for the next one I'm going to train with the larger file size and then try compressing it to smaller files in Kohya SS so I can compare the quality with an X/Y Plot in Automatic1111 and just upload two or more sizes as different versions. We'll see!
Hey @Honest_Abe! I just uploaded a new LoRA that's kind of an "All in One" because it covers a lot of the POV basics, so you could consider justifying the space because it has multiple concepts in one file, but if not I was also able to compress it like I'd mentioned previously and got a working version down to 74MB! So please feel free to check it out here and let me know what you think and what you create! Cheers! https://civitai.com/models/144203
@redcreamsoda Oh wow! Can't wait to try!